Student Performance Prediction Using Machine Learning

1.4 Significance of the study

The project has potential significance for several groups. Teachers may use prediction as supplementary evidence when deciding which students deserve closer attention. School administrators may use aggregated model outputs to support academic monitoring and resource planning. Students and parents may use the information to understand potential performance risk. Educational researchers may use the project as an example of an end-to-end supervised learning workflow.

1.5 Scope of the study

The study focuses on supervised classification for student performance prediction. The proposed input variables are age, gender, attendance, study hours, previous scores, sleep hours, tutoring, transportation, internet access, parent education, extracurricular activities, and family income when available. The proposed algorithms are Logistic Regression, Decision Tree, Random Forest, and SVM. The study covers data acquisition, cleaning, EDA, feature engineering, model development, tuning, evaluation, comparison, and integration into a simple prediction application.

The scope does not include causal inference, automatic educational decision-making, or a claim that the model can measure a student’s ability in a definitive manner. The final target definition must be selected before model training and used consistently throughout the experiment.

1.6 Limitations of the study

• The public dataset may not represent all schools, grade levels, or educational contexts.
• Some proposed variables may not be available in the final downloaded dataset.
• A categorical target may simplify a more complex continuous academic outcome.
• Class imbalance can make accuracy alone insufficient for model selection.
• Model performance depends on data quality, preprocessing, feature definitions, and evaluation strategy.
• Feature importance represents predictive association rather than causation.
• The current project materials do not contain completed empirical model measurements.

1.7 Target users and beneficiaries

II. Literature Review or Related Work

This section establishes the conceptual basis for predicting student performance with machine learning. The discussion follows the scope of the project presentation and focuses on the role of student-related features, classification, and comparison of several commonly used supervised learning algorithms.

2.1 Student Academic Performance

Student academic performance is a multidimensional outcome. It can reflect prior achievement, learning behavior, attendance, access to resources, personal circumstances, and support received from family or educational institutions. The project therefore does not treat performance as a result of a single variable. Instead, it proposes combining multiple student characteristics as input features for prediction.

2.2 Factors Affecting Student Performance

The proposed feature set includes both demographic and learning-related variables. Attendance and study hours represent learning engagement, while previous scores provide information about prior achievement. Sleep hours and extracurricular activities represent lifestyle and time allocation. Tutoring, internet access, parent education, transportation, and family income where available represent aspects of educational support and access to resources.

2.3 Machine Learning for Student Performance Prediction

Machine learning can be applied to educational datasets to discover patterns and construct predictive models. In a supervised learning setting, the algorithm learns a relationship between input features and a known target variable. For this project, the target is intended to represent student performance, with the presentation specifying outputs such as High, Medium, Low or Pass/Fail.

Comparing several algorithms is important because different models make different assumptions about the relationship between features and outcomes. The proposed study therefore compares Logistic Regression, Decision Tree, Random Forest, and Support Vector Machine.

2.4 Logistic Regression

Logistic Regression is a supervised classification method that estimates the probability of a categorical outcome from a set of input variables. It provides a relatively interpretable baseline and can be useful for understanding how changes in input features are associated with the predicted class. In this project, Logistic Regression provides a baseline against which the other proposed classifiers can be compared.

2.5 Decision Tree

A Decision Tree classifies observations through a sequence of decision rules based on feature values. Its tree structure can be relatively easy to interpret because a prediction can be traced through a series of splits. This makes it useful for an educational prediction system where understanding the decision process may be valuable.

2.6 Random Forest

Random Forest is an ensemble method that combines predictions from multiple decision trees. By aggregating many trees, it can model nonlinear relationships and interactions between variables while reducing reliance on a single tree. The proposed study includes Random Forest to test whether an ensemble approach can improve predictive performance.

2.7 Support Vector Machine

Support Vector Machine (SVM) is a supervised learning algorithm that seeks a decision boundary capable of separating classes. Depending on the selected kernel and preprocessing, SVM can model complex relationships between features. It is included to provide another classification approach for comparison.

2.8 Research Gap and Motivation

The project is motivated by the practical gap between evaluating performance after examinations and identifying students who may need assistance earlier. Rather than relying on a single indicator, the proposed system combines multiple student characteristics and compares several algorithms. The presentation does not claim that one algorithm is already superior; determining the best-performing model is itself one of the research questions.

2.9 Comparison of candidate algorithms

III. Materials and Methods or Methodology

The methodology follows a supervised machine learning workflow: data acquisition, exploratory analysis and cleaning, feature preparation, model training, model comparison, evaluation, and integration into a prediction interface. The project work plan assigns data acquisition and initial EDA to Week 1, feature engineering and baseline training to Week 2, model tuning and comparison to Week 3, interface construction to Week 4, testing to Week 5, and final presentation and submission to Week 6.

3.1 Research workflow

1. Define the prediction problem and final target variable.

2. Acquire and document the selected dataset.

3. Inspect data types, missing values, duplicates, invalid values, and class distribution.

4. Conduct exploratory data analysis.

5. Clean and preprocess the data.

6. Encode categorical variables and scale features where required.

7. Split the dataset into training and testing subsets.

8. Train Logistic Regression, Decision Tree, Random Forest, and SVM.

9. Tune hyperparameters using a validation strategy on training data.

10. Evaluate final models on the isolated test set.

11. Compare accuracy, precision, recall, and F1-score.

12. Analyze important predictive factors.

13. Save the final preprocessing pipeline and trained model.

14. Expose the model through FastAPI and connect the Vue.js frontend.

15. Test and document the complete system.

3.2 Dataset source and data collection plan

The project presentation proposes using a publicly available dataset from Kaggle and identifies the Student Performance Prediction Dataset as the suggested source. The paper therefore treats the Kaggle dataset as the planned data source. The final experimental paper should report the exact number of observations, class distribution, missing values, and final feature list after the dataset has been downloaded and inspected.

Suggested source identified in the project presentation: Kaggle Student Performance Prediction Dataset.

3.3 Features and Target Variable

The planned input features are age, gender, attendance, study hours, previous scores, sleep hours, tutoring, transportation, internet access, parent education, extracurricular activities, and family income if available. The presentation specifies that the prediction output may be High, Medium, Low or Pass/Fail. The final implementation should select one target definition before training and use the same definition consistently during preprocessing, training, and evaluation.

 

3.4 Data Preprocessing

Before model training, the dataset should be checked for missing values, duplicated observations, inconsistent category labels, and unsuitable data types. Numerical variables should be converted to appropriate numeric formats, while categorical variables should be encoded for algorithms that require numerical input. Where scaling is required, particularly for SVM and potentially Logistic Regression, scaling should be performed within a training pipeline to avoid data leakage.

The dataset should then be separated into training and testing subsets. The split must be performed before fitting preprocessing transformations so that information from the test set is not used during training.

3.5 Exploratory Data Analysis

Exploratory Data Analysis (EDA) will be used to understand distributions, relationships, class balance, and potential anomalies. Descriptive statistics and visualizations can be used to examine attendance, study hours, previous scores, and other variables in relation to the selected target. The project work plan places initial EDA and cleaning in Week 1.

3.6 Feature Engineering

Feature engineering will prepare the raw variables for machine learning. Categorical variables can be encoded using suitable techniques such as one-hot encoding, while numerical variables can be retained or scaled depending on the algorithm. Feature selection should be guided by the available dataset and the research question concerning which factors most strongly influence performance.

3.7 Proposed Machine Learning Models

 

3.8 Model Training and Tuning

A baseline version of each algorithm should first be trained using the prepared training data. The models can then be tuned using an appropriate validation strategy, such as cross-validation on the training set. Hyperparameters should be selected without using the final test set. The final test set should be used only for unbiased comparison after model development is complete.

3.9 Evaluation Metrics

For the proposed classification task, the project specifies accuracy, precision, recall, and F1-score as evaluation metrics.

 

The final paper should report the metric values for each model after training. Since the current project presentation contains proposed methodology rather than completed experimental measurements, this version does not invent numerical results.

3.10 Software, Frameworks, and Development tools

The planned system separates the prediction interface from the machine learning backend. Vue.js is proposed for the frontend, FastAPI for the backend, and Python with Scikit-learn for machine learning. A user enters student information through the frontend. The backend receives the input, applies the same preprocessing used during training, passes the processed data to the selected trained model, and returns the predicted performance category.

IV. System analysis and design

The proposed system transforms the machine-learning experiment into a practical prediction application. The design separates the user interface, prediction API, preprocessing pipeline, and trained model. This separation supports maintainability and allows the final model to be updated without redesigning the complete application.

4.2 Functional requirements

4.3 Non-Functional requirements

4.4 Proposed system architecture

4.5 Data flow

4.6 Machine Learning training pipeline

1. Load the final dataset.

2. Separate input features and target.

3. Split data into training and testing subsets.

4. Fit preprocessing on training data only.

5. Train four candidate classifiers.

6. Tune hyperparameters using the training data.

7. Evaluate final models on the isolated test set.

8. Select and save the final model and preprocessing pipeline.

4.7 Prediction API Design

The FastAPI service is intended to expose a prediction endpoint that accepts a structured representation of the selected student features. The endpoint should validate the request, apply the saved preprocessing pipeline, invoke the selected model, and return the predicted performance category. If probability estimates are exposed, they should be labeled as model probabilities rather than certainty.

4.8 User interface and User flow

The frontend should provide a student-information form corresponding to the final model features. After submission, the system should display the predicted category and, where implemented, a concise explanation of important predictive factors. The interface should avoid language suggesting that the prediction is a definitive judgment.

Dashboard (Before Upload Dataset)

Dashboard (After prediction)

V. Results and Evaluation

This section defines the results that should be reported after implementation. The project presentation specifies the expected system outputs but does not provide completed experimental values. Consequently, numerical accuracy, precision, recall, F1-score, confusion matrices, and feature-importance values are intentionally left for insertion after the models are trained and evaluated.

5.1 Dataset Results

The completed study should report the final number of records, the number of input features, target classes, missing-value counts, and the distribution of the target classes. These values should be taken directly from the downloaded and cleaned dataset.

5.2 Model Performance

 

Result Base on Model Selected

5.3 Model Comparison

The best-performing model should be selected based on the evaluation criteria defined for the final target formulation. Accuracy alone may not be sufficient when class distributions are imbalanced, so precision, recall, and F1-score should also be considered. The final comparison should explain not only which model has the highest score but also whether the difference is meaningful for the intended educational use.

5.4 Important Factors Affecting Performance

The final system is expected to show important factors affecting student performance. For tree-based models, feature importance can provide a direct ranking. For other models, coefficients or model-agnostic explanation methods may be considered. The interpretation must be based on the actual trained model and should not be presented as causal evidence.

5.5 Prediction Application

The planned application should allow users to enter student ID for prediction, and display the resulting performance category. It should also provide a comparison of model performance and show important factors associated with risk prediction. These functions correspond to the expected outputs defined in the project presentation.

VI. Discussion

The proposed study addresses a practical educational problem: identifying students who may be at risk before poor outcomes become evident through formal examinations. Its main strength is the combination of multiple categories of information rather than reliance on a single performance indicator. Attendance, study hours, previous scores, learning support, access to resources, and other characteristics can collectively provide a richer representation of student circumstances.

The comparison of four different algorithms is also important. Logistic Regression provides a simple baseline, Decision Tree provides an interpretable rule-based model, Random Forest provides an ensemble alternative, and SVM provides a margin-based approach. The comparison can reveal whether a relatively simple model is sufficient or whether a more complex model provides substantially better predictive performance.

However, prediction should not be interpreted as a definitive judgment about a student's ability. A model can identify patterns in historical data, but it cannot establish that a particular factor causes a student's performance. Educational decisions should therefore use predictions as supporting information alongside teacher judgment and direct engagement with students.

6.2 Limitations

Another limitation is that the project presentation does not yet provide completed experimental results. The final empirical conclusions must therefore be based on the actual dataset, preprocessing decisions, trained models, and evaluation results. If the final dataset differs from the planned source or if the target definition changes, the methodology and results should be updated accordingly.

6.3 Threats to validity

Internal validity may be threatened by leakage between training and testing data, repeated tuning against the test set, or inconsistent preprocessing. External validity may be limited if the dataset differs from the target educational population. Construct validity may be affected by the definition of performance categories. These threats can be reduced through an isolated test set, transparent target definitions, reproducible pipelines, and careful documentation.

6.4 Ethical considerations

A student prediction system can affect how learners are perceived. Ethical use therefore requires privacy protection, minimal data collection, transparency about model limitations, and human oversight. Predictions should support educational conversations rather than automatically determine opportunities, interventions, or labels.

6.5 Research contribution

The contribution of this project is an end-to-end educational prediction framework rather than a single algorithm. It connects a public student dataset, feature preparation, comparative classification, evaluation, model interpretation, and a deployable prediction interface. The project can therefore serve as a practical demonstration of how machine learning can move from data analysis to an application while maintaining attention to evaluation and responsible use.

VII. Conclusion and Future Work

This paper presents a proposed machine learning approach for predicting student academic performance using multiple demographic, academic, lifestyle, support, and access-related features. The project is motivated by the need for earlier identification of students who may require academic support. It proposes four classification algorithms—Logistic Regression, Decision Tree, Random Forest, and Support Vector Machine—and evaluates them using accuracy, precision, recall, and F1-score.

The project also extends beyond model training by proposing a simple prediction application using Vue.js, FastAPI, Python, and Scikit-learn. The planned system will allow users to enter student information, obtain a predicted performance category, compare algorithm performance, and examine important factors affecting prediction.

Future work should complete model training and tuning, validate the models on unseen data, investigate class imbalance, and compare performance across alternative target definitions. Additional work could include collecting richer longitudinal student data, integrating actual LMS engagement data rather than mock or unavailable variables, adding explainability features, and conducting user evaluation with teachers or educational administrators. These improvements could make the system more reliable and useful as an educational early-warning tool.

References

[1]     Kaggle. (n.d.). Student Performance Prediction Dataset. Kaggle. Dataset source identified in the Group 13 project presentation: https://www.kaggle.com/datasets/prajwalkanade/student-performance-prediction-dataset.

[2]     Scikit-learn Developers. (n.d.). Scikit-learn: Machine Learning in Python. User Guide and API documentation.

[3]    Breiman, L. (2001). Random Forests. Machine Learning, 45, 5–32. https://doi.org/10.1023/A:1010933404324.

[4]    Cortes, C., & Vapnik, V. (1995). Support-vector networks. Machine Learning, 20, 273–297. https://doi.org/10.1007/BF00994018.

[5]    Pedregosa, F., et al. (2011). Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research, 12, 2825–2830.

Appendix A. Proposed System Workflow

The proposed workflow is: (1) collect student data, (2) clean and preprocess the dataset, (3) perform exploratory analysis, (4) engineer and encode features, (5) split data into training and testing sets, (6) train Logistic Regression, Decision Tree, Random Forest, and SVM models, (7) evaluate models using accuracy, precision, recall, and F1-score, (8) select the most appropriate model, and (9) deploy the prediction model through a FastAPI backend and Vue.js frontend.

Appendix B. Feature Dictionary